This is all true - machine code needs to be "basically perfect" to work. However, there are lots of problems in CS that are easier to check the answer to a sol

This is all true - machine code needs to be "basically perfect" to work. However, there are lots of problems in CS that are easier to check the answer to a solution than to solve in the first place. It may turn out to be the case that a well-tuned model can quickly produce solutions to some code-generation problems, that those solutions have a high enough likelihood of being correct, that it's fast enough to check (and maybe try again), and that this entire process is faster than state-of-the-art classical algorithms. However, if that were the case, I might also expect us to be able to extract better algorithms from the model - intuitively, machine code generation "feels" like something that's just better implemented through classical algorithms. Have you met a human that can do register allocation faster than LLVM?